Anthropic Launches OSS Scanner for Automated Security Audits
Anthropic has released an opt-in security scanning service that uses AI to identify vulnerabilities in open-source software projects.
Anthropic's OSS Scanner service provides automated security auditing for open-source repositories.
- Anthropic has introduced an automated, opt-in service to scan open-source projects for security vulnerabilities.
- The system builds projects in an isolated, network-restricted environment to conduct automated analysis using AI models.
- Reports are sent directly to maintainers, providing reproduction steps and potential patches without human review.
- Participation is open to established projects with critical infrastructure impact, following a model similar to OSS-Fuzz.
Overview of the Service
Anthropic has officially launched OSS Scanner, an automated security service designed to assist open-source maintainers in identifying potential vulnerabilities within their repositories. The service is an opt-in program that utilizes Anthropic's AI models to conduct security audits, offering an alternative to traditional, slower coordinated vulnerability disclosure (CVD) processes that often rely on manual human review.
By leveraging its own internal research and experience from initiatives like Project Glasswing, Anthropic aims to provide maintainers with rapid feedback on their codebases. Unlike the company's standard CVD process, which involves human assessment and can be time-intensive, OSS Scanner allows projects to receive findings directly from AI agents as soon as the scans are complete.
How the Security Scanner Functions
The core of the OSS Scanner service is an automated pipeline that operates within an isolated virtual machine. To enroll, maintainers must submit a pull request to the project's GitHub repository. The integration requires a project.yaml configuration file and a Dockerfile that defines the build process. Once configured, the system executes the following steps:
- Initial Build: The project is built using a
Dockerfile, which runs with network access to download and install all necessary dependencies. - Isolated Scanning: Following the build, the environment is moved to an isolated segment with no internet access. This ensures that the subsequent security analysis occurs in a secure, controlled state.
- Automated Analysis: AI agents analyze the source code and built environment to detect potential flaws. These agents are tasked with verifying bugs, performing root cause analysis, and where possible, suggesting specific patches.
- Reporting: Findings are emailed directly to a primary contact address provided by the project maintainers. The reports include detailed reproduction steps to facilitate quick remediation.
Anthropic has emphasized that the reports generated by this service are not reviewed by humans. Consequently, these findings are not subject to standard disclosure waiting periods, and the reports are kept private between Anthropic and the project maintainers.
Eligibility and Enrollment
The program is currently targeted at established open-source projects that play a critical role in infrastructure or user security. Anthropic notes that it follows criteria similar to those used by OSS-Fuzz, assessing projects based on factors such as their exposure to remote attacks and the number of downstream dependencies. Each enrollment request is reviewed manually to confirm that the person submitting the request is a core maintainer of the project.
Maintainers have control over the scanning process through their project.yaml files. This includes the ability to specify email contacts, provide GPG keys for encrypted reporting, and include an optional threat_model.md file. The threat model document is particularly useful for providing context to the scanner, such as defining severity rubrics, identifying untrusted input points, or specifying which components should be prioritized during an audit.
Security and Operational Considerations
Anthropic provides tools for maintainers to validate their configurations before submitting them. The tools/validate.py script checks the project.yaml file for errors, while the tools/check utility allows maintainers to simulate the build environment locally. This helps ensure that the project will build successfully within Anthropic's infrastructure.
The company notes that the tools/check utility should be used with caution, as it runs build processes locally. While the use of QEMU containers provides a degree of isolation, it can still interact with local host services. Maintainers are advised to test only trusted code. As the service continues to evolve, Anthropic intends to refine the process, with potential updates to reporting formats and expanded support for maintainers who are already managing high volumes of security reports.
Enjoyed this?
Get more posts like this delivered to your inbox.
